Learning device, learning method and recording medium

Abstract

The dataset supply unit supplies a learning dataset. The recognition unit outputs the recognition result for the recognition object data in the supplied learning dataset. Further, the intersection matrix computation unit computes the intersection matrix based on the learning dataset. The recognition loss computation unit computes the recognition loss using the recognition result, the intersection matrix, and the correct answer data given to the recognition object data. Then, the updating unit updates the parameters of the recognition unit based on the recognition loss.

What is claimed is: 
     
       1. A learning device for performing learning using a learning dataset,
 wherein the learning dataset includes a plurality of partial datasets to which at least a part of all categories of recognition objects is assigned as a responsible range, 
 wherein all categories of the recognition objects are assigned to one of the plurality of partial datasets, 
 wherein each of recognition object data included in the partial dataset is given correct answer data indicating any one of the categories belonging to the responsible range of the partial dataset, or indicating that the category of the recognition object does not belong to the responsible range of the partial dataset, and 
 wherein the learning device comprises: 
 a memory configured to store instructions; and 
 one or more processors configured to execute the instructions to: 
 supply the learning dataset; 
 output a recognition result for the recognition object data in the supplied learning dataset; 
 compute an intersection matrix based on the learning dataset; 
 compute the recognition loss using the recognition result, the intersection matrix, and the correct answer data given to the recognition object data; and 
 update parameters of the recognition unit based on the recognition loss. 
 
     
     
       2. The learning device according to  claim 1 , wherein the one or more processors compute the intersection matrix based on a first prior distribution that is a rate that the recognition object data in the learning dataset belongs to each partial dataset, a second prior distribution that is a rate of each category included in the partial dataset, and a code indicating the responsible range assigned to each of the partial datasets. 
     
     
       3. The learning device according to  claim 2 , wherein the one or more processors are further configured to execute the instructions to:
 estimate the first prior distribution from the learning dataset; and 
 estimate the second prior distribution from the learning dataset. 
 
     
     
       4. The learning device according to  claim 2 , wherein the the one or more processors compute a transition matrix using the first prior distribution, the second prior distribution, and the code indicating the responsible range, and compute the intersection matrix using an inverse matrix of the transition matrix. 
     
     
       5. The learning device according to  claim 1 , wherein the the one or more processors compute the recognition loss by weighting and adding losses between the recognition result for all the recognition object data included in the learning dataset and all the categories of the recognition objects using elements of the intersection matrix as weights. 
     
     
       6. A learning method using a learning dataset,
 the learning dataset including a plurality of partial datasets to which at least a part of all categories of recognition objects is assigned as a responsible range, 
 all categories of the recognition objects being assigned to one of the plurality of partial datasets, 
 each of recognition object data included in the partial dataset being given correct answer data indicating any one of the categories belonging to the responsible range of the partial dataset, or indicating that the category of the recognition object does not belong to the responsible range of the partial dataset, 
 the learning method comprises: 
 supplying the learning dataset; 
 outputting a recognition result for the recognition object data in the supplied learning dataset; 
 computing an intersection matrix based on the learning dataset; 
 computing the recognition loss using the recognition result, the intersection matrix, and the correct answer data given to the recognition object data; and 
 updating parameters of the recognition unit based on the recognition loss. 
 
     
     
       7. A non-transitory computer-readable recording medium for recording a program for a learning process using a learning dataset,
 the learning dataset including a plurality of partial datasets to which at least a part of all categories of recognition objects is assigned as a responsible range, 
 all categories of the recognition objects being assigned to one of the plurality of partial datasets, 
 each of recognition object data included in the partial dataset being given correct answer data indicating any one of the categories belonging to the responsible range of the partial dataset, or indicating that the category of the recognition object does not belong to the responsible range of the partial dataset, 
 the learning process comprises: 
 supplying the learning dataset; 
 outputting a recognition result for the recognition object data in the supplied learning dataset; 
 computing an intersection matrix based on the learning dataset; 
 computing the recognition loss using the recognition result, the intersection matrix, and the correct answer data given to the recognition object data; and 
 updating parameters of the recognition unit based on the recognition loss.